| --- |
| license: apache-2.0 |
| pretty_name: TriageBench |
| language: |
| - en |
| tags: |
| - clinical-ai |
| - llm-security |
| - ai-agents |
| - healthcare |
| - haarf |
| - quokkaguard |
| - synthetic |
| size_categories: |
| - n<1K |
| task_categories: |
| - other |
| --- |
| |
| # TriageBench |
|
|
| **Risk-calibrated escalation routing that sends only genuinely ambiguous agent actions to clinicians.** |
|
|
| TriageBench is the benchmark corpus for *Calibrating Human Oversight for Clinical AI Agents: Routing Only the Ambiguous Tail to Clinicians*, the `qutriage` module of the QUOKKAGUARD program. It ships with the [`qutriage`](https://github.com/quome-cloud/qutriage) repository, which contains the `qfire` gateway layer under test, the experiment harness, and the paper. |
|
|
| 600 synthetic clinical agent actions in four categories (easy-legit, easy-attack, stealth-attack with gray-zone signals, suspicious-legit with elevated signals), each with three simulated-expert labels, a severity, and a time-to-care impact, plus the fitted calibration (calibration_fitted.json); fully tracked, regenerable at seed 42 with scripts/009-hitl-calibration/gen_triage.py. |
|
|
| **All data are synthetic.** No real patient data or protected health information (PHI) is included; clinical content is generated from templates with fixed seeds. |
|
|
| ## Files |
|
|
| | File | Size | Rows | |
| |---|---|---| |
| | `calibration_fitted.json` | 434 B | | |
| | `triagebench/actions.jsonl` | 363 KB | 600 rows | |
| | `triagebench/raters.jsonl` | 93 B | 3 rows | |
|
|
| ### Record schemas |
|
|
| - `triagebench/actions.jsonl`: `action`, `category`, `expert_labels`, `gold_decision`, `id`, `legit`, `signals`, `source_bench`, `tool`, `tool_risk`, `true_harm` |
| - `triagebench/raters.jsonl`: `noise`, `rater` |
|
|
| ## How it was generated |
|
|
| The corpus is produced by the generator in the paper repository. |
|
|
| ## Intended use |
|
|
| Evaluating the **oversight** enforcement layer of a clinical-agent security gateway (HAARF control C4): Every upstream guardrail in a clinical AI gateway answers a binary allow/block question, and strict enforcement over-blocks: the policy control alone false-denies 28.3% of legitimate actions, which is a real care-delay hazard and a source of alarm fatigue. |
|
|
| The experiments that consume it (E-series in the paper) are reproduced from the repository: |
|
|
| ```bash |
| git clone https://github.com/quome-cloud/qutriage |
| cd qutriage |
| cargo build --release |
| ``` |
|
|
| then follow the *Reproduce the experiments* section of its README. |
|
|
| ## Citation |
|
|
| This benchmark was built to evaluate a control of the **Healthcare AI Agents Regulatory Framework (HAARF)**, the source framework for the QUOKKAGUARD program. Please cite both the paper and HAARF: |
|
|
| ```bibtex |
| @unpublished{schwoebel2026qutriage, |
| author = {Schwoebel, James}, |
| title = {Calibrating Human Oversight for Clinical AI Agents: Routing Only the Ambiguous Tail to Clinicians}, |
| note = {Preprint. Quome, QUOKKAGUARD program (qutriage module)}, |
| year = {2026}, |
| url = {https://github.com/quome-cloud/qutriage} |
| } |
| |
| @unpublished{schwoebel2026haarf, |
| author = {Schwoebel, Jim and Frasch, Martin and Spalding, Art and Sewell, Ed and Englert, Phil and Halpert, Ben and Overbay, Collin and Semenec, Ingrida and Shor, Joel}, |
| title = {{HAARF}: Healthcare {AI} agents regulatory framework --- a comprehensive security verification standard for autonomous {AI} systems in clinical environments}, |
| note = {medRxiv Preprint}, |
| year = {2026}, |
| month = {April}, |
| doi = {10.64898/2026.04.09.26350519}, |
| url = {https://www.medrxiv.org/content/10.64898/2026.04.09.26350519v1} |
| } |
| ``` |
|
|
| ## License |
|
|
| Apache License 2.0. Copyright (c) 2026 Quome, Inc. |
|
|
